Broad-coverage model of prediction in human sentence processing

Abstract

The aim of this thesis is to design and implement a cognitively plausible theory of sentence processing which incorporates a mechanism for modeling a prediction and verification process in human language understanding, and to evaluate the validity of this model on specific psycholinguistic phenomena as well as on broad-coverage, naturally occurring text. Modeling prediction is a timely and relevant contribution to the field because recent experimental evidence suggests that humans predict upcoming structure or lexemes during sentence processing. However, none of the current sentence processing theories capture prediction explicitly. This thesis proposes a novel model of incremental sentence processing that offers an explicit prediction and verification mechanism. In evaluating the proposed model, this thesis also makes a methodological contribution. The design and evaluation of current sentence processing theories are usually based exclusively on experimental results from individual psycholinguistic experiments on specific linguistic structures. However, a theory of language processing in humans should not only work in an experimentally designed environment, but should also have explanatory power for naturally occurring language. This thesis first shows that the Dundee corpus, an eye-tracking corpus of newspaper text, constitutes a valuable additional resource for testing sentence processing theories. I demonstrate that a benchmark processing effect (the subject/object relative clause asymmetry) can be detected in this data set (Chapter 4). I then evaluate two existing theories of sentence processing, Surprisal and Dependency Locality Theory (DLT), on the full Dundee corpus. This constitutes the first broad-coverage comparison of sentence processing theories on naturalistic text. I find that both theories can explain some of the variance in the eye-movement data, and that they capture different aspects of sentence processing (Chapter 5). In Chapter 6, I propose a new theory of sentence processing, which explicitly models prediction and verification processes, and aims to unify the complementary aspects of Surprisal and DLT. The proposed theory implements key cognitive concepts such as incrementality, full connectedness, and memory decay. The underlying grammar formalism is a strictly incremental version of Tree-adjoining Grammar (TAG), Psycholinguistically motivated TAG (PLTAG), which is introduced in Chapter 7. I then describe how the Penn Treebank can be converted into PLTAG format and define an incremental, fully connected broad-coverage parsing algorithm with associated probability model for PLTAG. Evaluation of the PLTAG model shows that it achieves the broad coverage required for testing a psycholinguistic theory on naturalistic data. On the standardized Penn Treebank test set, it approaches the performance of incremental TAG parsers without prediction (Chapter 8). Chapter 9 evaluates the psycholinguistic aspects of the proposed theory by testing it both on a on a selection of established sentence processing phenomena and on the Dundee eye-tracking corpus. The proposed theory can account for a larger range of psycholinguistic case studies than previous theories, and is a significant positive predictor of reading times on broad-coverage text. I show that it can explain a larger proportion of the variance in reading times than either DLT integration cost or Surprisal.